How to integrate machine learning in solar cell manufacturing

Scientists in Korea have developed a new methodology to employ machine-learning models in "smart" solar cell manufacturing. They utilized data collected from equipment that closely resembles actual industrial manufacturing tools.

Researchers at Korea University have developed a machine learning model for predicting sheet resistance in phosphorus oxychloride (POCl3) doping processes in solar cell manufacturing.

“Our study aims to propose a methodology for integrating machine learning into industrial processes, with the goal of accelerating the advancement of Industry 4.0 and paving the way toward Industry 5.0,” the research's lead author, Seungtae Lee, told pv magazine.

“We utilized data collected from equipment that closely resembles actual industrial manufacturing tools,” he went on to say. “Using this data, we developed a machine learning model not only to predict the sheet resistance based on process conditions but also to optimize those conditions through Bayesian optimization to meet specific sheet resistance targets.”

In solar cell production, POCl3 is utilized as a liquid dopant precursor to create n-type layers in the thermal diffusion process.

For their modeling, the scientists considered different furnace process conditions and sheet resistance values. They collected 3,420 experimental data points, with 10 process variables being used as input parameters: Pre-deposition temperature; pre-deposition time; drive-in conditions; drive-in temperature; drive-in time; process gas parameters; POCl3 flow rate, O2 flow rate, and process pressure; wafer boat position; wafer slot number; and wafer position.

The research group used the SHapley Additive exPlanations (SHAP) method, which is a game-theory approach to explain the output of any machine learning model, to analyze the impact of each feature on sheet resistance prediction. “SHAP is an interpretability technique based on Shapley values from game theory,” it stressed. “It provides a comprehensive quantitative analysis that includes feature importance, the influence of each feature on model predictions, and the specific contributions of individual features to each prediction at the data-point level.”

The academics also used Bayesian optimization, which is commonly used to solve complex optimization problems by approximating an unknown objective function and efficiently identifying its minimum or maximum values, to identify the optimal process conditions by leveraging the trained machine-learning model. In particular, they sought to  identify the conditions that yield a sheet resistance close to 150 Ω/sq under “realistic” solar cell production conditions.

The team conducted 100 trials in the initial random sampling phase and 100 trials in the Bayesian optimization phase.

The proposed approach was found to achieve a more efficient and rapid optimization of process conditions compared to conventional and expensive trial-and-error methods used in the PV industry.

“We found that the model's learned representations and predictions are consistent with established physical and theoretical understanding. This provides confidence in the reliability and interpretability of the model in real-world manufacturing environments,” Lee further explained. “We believe that this methodology could be extended beyond solar cell manufacturing to a wide range of industrial processes.”

The proposed approach was described in the study “Bayesian-optimization-based approach for sheet-resistance control in silicon wafers toward automated solar-cell manufacturing,” published in Materials Science in Semiconductor Processing.